Patch notesUpdates

0.3.10: Free Models & Code Sign-In

Free models in Chat don't use your plan, up to 250 messages an hour with Zero Data Retention off, and a signed-out coding program can sign in from a chat.

Scroll down

AI Integrator Chat now offers free models, whose replies don't count toward your plan's usage (paid tools such as web search still do), up to 250 messages an hour. Chat asks before the first one, because for those replies Zero Data Retention and no-training are off and the model may run on any host. In Code, when a coding program such as Claude Code or Codex is signed out, you can sign it in from the chat that needs it, on your computer, your phone or the web.

Free models

Free models sit under Free, first in the model list, on the desktop, the web and iPhone. They are models that the companies serving them offer at no cost, and a reply from one does not count toward your plan, apart from paid tools such as web search. You can send up to 250 free messages an hour.

The first time you pick one, Chat asks you to turn free models on. Your other models run only on hosts we approved in the United States or Europe. A free model can run on any host that serves it, Zero Data Retention and no-training are off for its replies, and the company running it may keep your messages and train on them. Free models can also change or stop working without notice. The rest of your account keeps its privacy settings. Free models can be turned off again on the Privacy page in Settings, and temporary chats never offer them. Free AI models in Chat has the details.

Code sign-in

When a Code chat's coding program is signed out, the chat offers Sign in. The program's own sign-in runs on your computer, and the card shows what it is waiting for: a page to finish in your browser, a code to enter on the vendor's site, a code to paste back, or a key typed into a hidden field. When it finishes, Send again sends your last message; nothing is sent until you press it.

You can also sign in from your phone or the web for a chat on one of your computers, as long as that device can approve commands on it. There the sign-in shows as a row under the reply that failed, and it opens a sheet with the same steps. Steps only the computer can answer, such as typing an API key, ask you to finish there.

Other changes

  • Continue on moves a Code chat to another computer with its whole conversation. A chat that is too long or too large to move stops with a message before anything is sent.
  • A step in a plan can hold a short detail. Open the plan and click the arrow on that step to read it.
  • A subagent runs on Auto unless it was started read-only. When it asks to do something beyond its project, the Auto reviewer answers first, then the agent that started it, which asks you only when it is unsure. A request with no answer in 10 minutes is declined. Subagents, plans and questions in Code explains how a subagent on Auto gets its answers.
  • Tabs in the work pane share one width, fade at the ends, and show their close button when you point at them.
  • A code file shows its line numbers in a column beside the code and marks the line the cursor is on. Click a number to select that line.
  • While a reply is being written, the blinking cursor sits right after its last word, and a table appears a whole row at a time.
  • On iPhone, a button in a Code chat's header opens a live view of that chat's computer, and that view no longer pauses.
  • The phone's chat list stays complete, and comes back faster after the phone sleeps.
  • On the desktop, the title bar blurs along with the screen behind Settings, and File, Edit, View and the window buttons still work.

The release before this one, 0.3.9, pinned a coding agent's plan at the top of the chat and added question cards.